dunigan@MSR.EPM.ORNL.GOV (Tom Dunigan 576-2522) (06/04/91)
What kind of experiences (reliability, installability, performance ...) have folks had with the FDDI interfaces for Sun. This would include Sun FDDI interface, CMC's, others? thanks tom dunigan@msr.epm.ornl.gov
liam@dcs.qmw.ac.uk (William Roberts;) (06/18/91)
In <9106041306.AA09116@msr.EPM.ORNL.GOV> dunigan@MSR.EPM.ORNL.GOV (Tom Dunigan 576-2522) writes: >What kind of experiences (reliability, installability, performance ...) >have folks had with the FDDI interfaces for Sun. >This would include Sun FDDI interface, CMC's, others? The Sun FDDI card basically works. The main interesting things about it are: 1) It breaks the ring when the card is configured down (even if the Sun is still powered up). Use it single attached via a concentrator or with an optical bypass. 2) It didn't conform to RFC 1188 - puts out ARP requests with hardware type code 6 that give problems in a bridged environment. This can (now) be tweaked at kernel build time. 3) It gives *no measurable performance increase* for NFS - even with both client and server on FDDI and tests which don't involve server disk access there is no increase in performance over Ethernet connections. We have a pair of the CMC cards but we haven't tested them yet. Mail me a reminder in a month's time if you don't get any alternative responses. -- % William Roberts Internet: liam@dcs.qmw.ac.uk % Queen Mary & Westfield College UUCP: liam@qmw-dcs.UUCP % Mile End Road Telephone: +44 71 975 5234 % LONDON, E1 4NS, UK Fax: +44 81-980 6533
liam@dcs.qmw.ac.uk (William Roberts;) (06/21/91)
In <3384@redstar.dcs.qmw.ac.uk> liam@dcs.qmw.ac.uk (William Roberts;) writes: >In <9106041306.AA09116@msr.EPM.ORNL.GOV> dunigan@MSR.EPM.ORNL.GOV (Tom Dunigan >576-2522) writes: >3) It gives *no measurable performance increase* for NFS - even with both >client and server on FDDI and tests which don't involve server disk access >there is no increase in performance over Ethernet connections. I had some mail saying: | I'd be *fascinated* to read or hear more about your experiences |with FDDI cards, especially Sun's and *especially* why Sun's doesn't |get any speed inprovements with FDDI vs. Ethernet! I'm quite prepared to believe that Sun4 machines (I was using a Sun 4/280 and a Sun 4/160) can drive the FDDI network at greater than Ethernet bandwidth: one of the planned experiments was to change the FDDI driver software so that it generated continuous "idle traffic" transmitting strips of a VME framebuffer, sending to another machine which writes such things direct into its display memory. What my experiments showed was that there is no performance improvement for NFS operations benchmarked with the Legato "nhfsstone" benchmark: I've a paper on this in the UK Sun User Group conference in September (in fact an extention of a paper I gave at the European Unix User Group Autumn '90 conference) and that's probably the best way to look at what I've done. There should be a report on the overall FDDI Pilot project (it's being done for the JNT: the body which runs the UK academic network) but I don't know the timescale on that any more. FDDI doesn't make any difference because present machines can't really compute at those speeds (well, Crays and the like perhaps, but nothing that Sun manufacture). The only applications which ship data around that fast and which might benefit from FDDI are ones where the machines at either end don't have to manipulate the data as it comes in - they just pump bursts of data for later use. Obvious examples are live video and very low-level data transfer between peripherals (FDDI started life as something for IBM channels, I believe): the only one which seems pertinent to workstations is swapping over FDDI. -- % William Roberts Internet: liam@dcs.qmw.ac.uk % Queen Mary & Westfield College UUCP: liam@qmw-dcs.UUCP % Mile End Road Telephone: +44 71 975 5234 % LONDON, E1 4NS, UK Fax: +44 81-980 6533
barmar@think.com (Barry Margolin) (06/22/91)
In article <3393@redstar.dcs.qmw.ac.uk> liam@dcs.qmw.ac.uk (William Roberts;) writes: >I'm quite prepared to believe that Sun4 machines (I was using a Sun 4/280 and >a Sun 4/160) can drive the FDDI network at greater than Ethernet bandwidth: >one of the planned experiments was to change the FDDI driver software so that >it generated continuous "idle traffic" transmitting strips of a VME >framebuffer, sending to another machine which writes such things direct into >its display memory. Someone here did some tests with a program that just sends UDP or TCP data, not NFS, 4/3xx and 4/4xx machines, over FDDI. For certain datagram sizes he was able to get much better than Ethernet bandwidth. But at other sizes, it wasn't much better than Ethernet; I suspected that this was due to kernel buffering strategies optimized for Ethernet's smaller MTU. >FDDI doesn't make any difference because present machines can't really compute >at those speeds (well, Crays and the like perhaps, but nothing that Sun >manufacture). Be aware that the machines you said you tried are not "present machines." The 4/280 and 4/160 have been around for almost four years. The current top-of-the-line Sun-4's are about an order of magnitude faster. -- Barry Margolin, Thinking Machines Corp. barmar@think.com {uunet,harvard}!think!barmar